La Agente 'Optima: Toward Agentic Self-Driving Laboratories

Summary

Self-driving laboratories combine automated experiments with adaptive decision-making, but their campaigns often require specialists to translate scientific goals into executable closed-loop procedures. La Agente 'Optima is an agentic framework that constructs and supervises Bayesian optimization campaigns across computational and experimental systems while preserving a persistent optimization state. It separates LLM reasoning from campaign execution, allowing repetitive loops to run consistently and returning control to the agent when interpretation or revision is needed; decisions remain auditable. The authors evaluated it through ablations, five digital discovery tasks, and two physical platforms, finding that campaigns stayed executable as both the scientific problem and execution environment changed. In a contact-angle campaign, the system detected and corrected a mid-run measurement failure, moving the result from 71.4 to 67.8 degrees, just above the 64-66 degree target range. It inferred that the target was probably unattainable with the available reagents and recommended changing the formulation. In a five-day multi-objective flow-chemistry campaign, 23 experiments increased yield from 30% to 59%. Despite substantial inference costs, the agent used less money and starting material than a human-directed campaign and selected a more mass-efficient operating point.